Gentoo Forums
Gentoo Forums
Gentoo Forums
Quick Search: in
Acovea analysis results against real world programs
View unanswered posts
View posts from last 24 hours

Goto page Previous  1, 2  
Reply to topic    Gentoo Forums Forum Index Portage & Programming
View previous topic :: View next topic  
Author Message
procyon112
n00b
n00b


Joined: 28 Apr 2005
Posts: 16
Location: Seattle, Washington, USA

PostPosted: Sat Apr 30, 2005 5:58 am    Post subject: Reply with quote

Quote:
This is not too different from now, except for step 3


My contention is that it is step 2 that is significantly different. You want to evolve individuals that operate over a wider group of programs, so each individual should be tested against each benchmark and given a fitness score with a large penalty for breaking a benchmark. I see your point in the dangers of overoptimising for the domain, but I also question if your method of evolving seperately for each benchmark doesn't overoptimize for a domain that is so specific as to be meaningless. By filtering out the flag combinations that are "not nice", yet testing in aggregate, while it is true that the domain is still too specific to be taken as an absolute optimization for a real world situation, at least you are drawing out flag combinations that are mutually beneficial over a wider domain than a single benchmark.

I have read the original thread and it seems that the evidence of improved stability with the evolved set of flags seems to be more anecdotal, whereas this particular thread points out real stability problems compared with -O2. Since this particular case of MP3 encoding breaks frequently on aggressive optimization AND shows a decent benchmark, it might be very beneficial to add it, and any other known "breaking" benchmarks to the list so that we even further increase the domain space and end up with "nicer" flags. By adding test cases known to break, you can reduce the problem of strongly correlated flags prone to breaking apps and amplify the effect of strongly correlated flags that improve performance. This would make the flags more trustworthy for safety, and be able to more accurate prescribe a set of flags that the user should not "pick and choose" from for optimal stability AND performance.

My main beef is with the testing each benchmark seperately though, as I think that dilutes the data you are looking for.
Back to top
View user's profile Send private message
Hypnos
Advocate
Advocate


Joined: 18 Jul 2002
Posts: 2869
Location: Omnipresent

PostPosted: Sat Apr 30, 2005 7:02 am    Post subject: Reply with quote

There's probably no harm in mashing the codes together if the population is large enough, to capture the diversity of "friendly" flags; there would be technical difficulties in weighting the codes properly (not all take the same time to run, etc.).

A greater variety of codes would help, of course in relation to the real world. A wavelet code might be a nice addition.

Finally, the newest major version of Acovea (released a few days ago) can test on return value, not just stdout -- this might be useful for validating scientific codes that include test harnesses.
_________________
Personal overlay | Simple backup scheme
Back to top
View user's profile Send private message
Display posts from previous:   
Reply to topic    Gentoo Forums Forum Index Portage & Programming All times are GMT
Goto page Previous  1, 2
Page 2 of 2

 
Jump to:  
You cannot post new topics in this forum
You cannot reply to topics in this forum
You cannot edit your posts in this forum
You cannot delete your posts in this forum
You cannot vote in polls in this forum